Spatio-temporal data fusion for adversarially resilient graph neural networks in construction progress management

The digital transformation in the construction industry has catalyzed the adoption of machine learning (ML) systems for decision-making processes. However, this reliance raises critical concerns about the security and resilience of ML models against adversarial threats. This study emphasizes a data-centric perspective, fostering the integration of spatio-temporal data fusion to improve the robustness of ML systems. Spatio-temporal GNNs (RNN-GCN, GCN-RNN) achieved superior accuracy compared to spatial models (GCN, GAT), with RNN-GCN reaching 81% and GCN-RNN 75%, versus 71% for both GCN and GAT. Under adversarial testing, the models displayed different susceptibilities. Label flipping modestly reduced performance (RNN-GCN to 73%, GCN-RNN to 69%), while random label assignment moderately affected all models. Feature perturbation produced minimal adverse effects. However, FGSM emerged as the most aggressive attack, drastically lowering RNN-GCN to 57% and GCN-RNN to 26%, revealing temporal models’ inherent vulnerability to gradient-targeted attacks. While spatial information aggregation could reduce the damage from poisoned records, temporal dependencies further enhanced the models’ resilience against attacks. These findings are significant because decisions derived from ML-based decision support systems have considerable financial impact, emphasizing the necessity to ensure that data perturbations do not compromise model performance.